Message-ID: <4E9C11F5-5197-4DFC-900C-D6831D5441D6@oracle.com>
Date: Tue, 29 Oct 2024 13:09:40 +0000
From: Chuck Lever III <chuck.lever@...cle.com>
To: Jeff Layton <jlayton@...nel.org>
CC: Neil Brown <neilb@...e.de>, Dai Ngo <dai.ngo@...cle.com>,
        Tom Talpey
	<tom@...pey.com>, Olga Kornievskaia <okorniev@...hat.com>,
        Linux NFS Mailing
 List <linux-nfs@...r.kernel.org>,
        "linux-kernel@...r.kernel.org"
	<linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 2/2] nfsd: allow for more callback session slots



> On Oct 29, 2024, at 6:28 AM, Jeff Layton <jlayton@...nel.org> wrote:
> 
> On Mon, 2024-10-28 at 15:05 -0400, Chuck Lever wrote:
>> On Mon, Oct 28, 2024 at 10:26:27AM -0400, Jeff Layton wrote:
>>> nfsd currently only uses a single slot in the callback channel, which is
>>> proving to be a bottleneck in some cases. Widen the callback channel to
>>> a max of 32 slots (subject to the client's target_maxreqs value).
>>> 
>>> Change the cb_holds_slot boolean to an integer that tracks the current
>>> slot number (with -1 meaning "unassigned").  Move the callback slot
>>> tracking info into the session. Add a new u32 that acts as a bitmap to
>>> track which slots are in use, and a u32 to track the latest callback
>>> target_slotid that the client reports. While they are part of the
>>> session, the fields are protected by the cl_lock.
>>> 
>>> Fix nfsd41_cb_get_slot to always search for the lowest slotid (using
>>> ffs()), and change it to continually retry until there is a slot
>>> available.
>>> 
>>> Finally, convert the session->se_cb_seq_nr field into an array of
>>> counters and add the necessary handling to ensure that the seqids get
>>> reset at the appropriate times.
>>> 
>>> Signed-off-by: Jeff Layton <jlayton@...nel.org>
>>> ---
>>> fs/nfsd/nfs4callback.c | 107 +++++++++++++++++++++++++++++++++++--------------
>>> fs/nfsd/nfs4state.c    |   7 +++-
>>> fs/nfsd/state.h        |  12 +++---
>>> fs/nfsd/trace.h        |   2 +-
>>> 4 files changed, 89 insertions(+), 39 deletions(-)
>>> 
>>> diff --git a/fs/nfsd/nfs4callback.c b/fs/nfsd/nfs4callback.c
>>> index e38fa834b3d91333acf1425eb14c644e5d5f2601..64b85b164125b244494f9805840a0d8a1ccb4c1b 100644
>>> --- a/fs/nfsd/nfs4callback.c
>>> +++ b/fs/nfsd/nfs4callback.c
>>> @@ -406,6 +406,16 @@ encode_cb_getattr4args(struct xdr_stream *xdr, struct nfs4_cb_compound_hdr *hdr,
>>> hdr->nops++;
>>> }
>>> 
>>> +static u32 highest_unset_index(u32 word)
>>> +{
>>> + int i;
>>> +
>>> + for (i = sizeof(word) * 8 - 1; i > 0; --i)
>>> + if (!(word & BIT(i)))
>>> + return i;
>>> + return 0;
>>> +}
>>> +
>> 
>> Isn't this the same as ffz() or "ffs(~(x))" ?
>> 
> 
> No. I need the _last_ cleared bit in the word. It looks though like
> there is a fls() function that might be usable, so I could maybe do
> fls(~(x)). I'll experiment with that.
> 
>> 
>>> /*
>>>  * CB_SEQUENCE4args
>>>  *
>>> @@ -432,15 +442,38 @@ static void encode_cb_sequence4args(struct xdr_stream *xdr,
>>> encode_sessionid4(xdr, session);
>>> 
>>> p = xdr_reserve_space(xdr, 4 + 4 + 4 + 4 + 4);
>>> - *p++ = cpu_to_be32(session->se_cb_seq_nr); /* csa_sequenceid */
>>> - *p++ = xdr_zero; /* csa_slotid */
>>> - *p++ = xdr_zero; /* csa_highest_slotid */
>>> + *p++ = cpu_to_be32(session->se_cb_seq_nr[cb->cb_held_slot]); /* csa_sequenceid */
>>> + *p++ = cpu_to_be32(cb->cb_held_slot); /* csa_slotid */
>>> + *p++ = cpu_to_be32(max(highest_unset_index(session->se_cb_slot_avail),
>>> +        session->se_cb_highest_slot)); /* csa_highest_slotid */
>> 
>> This encoder doesn't hold the client's cl_lock, but does reference
>> fields that are updated while that lock is held.
>> 
>> Also, should a per-session spin lock be used instead of cl_lock?
>> 
>> Is locking even necessary? All of these calls are prepared in a
>> single-threaded workqueue. Reply processing can only free slots, so
>> it doesn't need to exclude session slot selection.
>> 
> 
> NFSv4 callbacks use async RPCs. The encoding and transmit is done in
> the context of the (single-threaded) workqueue, but the reply is
> handled by rpciod. Since we release the slot after the reply in rpciod
> context, I think we do require locking here.

Fair enough, there are some shared data items that have RMW
accesses in both the Call and Reply path.


> I also worry a bit that
> relying on the single-threaded workqueue may end up being a bottleneck
> at some point in the future, so I see having clear locking here as a
> benefit.
> 
> I'm open to switching to a per-session lock of some sort, but I don't
> see a real need here. Only one session will be used as the backchannel
> at a time, so there shouldn't be competing access between different
> sessions for the cl_lock. We are competing with the other uses of the
> cl_lock, but this one should be pretty quick. My preference would be to
> add extra locking only once it becomes clear that it's necessary.

I agree that some day we want to escape from the single-thread
workqueue paradigm.

I don't think this usage will generate a lot of contention now,
but it will pull the lock's cache lines away from the hotter
users on occasion.

Also, I think a per-session lock is more self-documenting to
show precisely what fields are being protected. I've often
had to look at breaking up a larger lock and found it hard
to figure out exactly what needs protection and why.

Thus I like per-session locking a little better, with a brief
comment that explains the locking design. (It's not a strong
preference, but I think we will be better served going forward).


>>> *p++ = xdr_zero; /* csa_cachethis */
>>> xdr_encode_empty_array(p); /* csa_referring_call_lists */
>>> 
>>> hdr->nops++;
>>> }
>>> 
>>> +static void update_cb_slot_table(struct nfsd4_session *ses, u32 highest)
>> 
>> Nit: Can you use the name "target" instead of "highest" ?
>> 
> 
> Will do.
> 
>> 
>>> +{
>>> + /* No need to do anything if nothing changed */
>>> + if (likely(highest == READ_ONCE(ses->se_cb_highest_slot)))
>>> + return;
>>> +
>>> + spin_lock(&ses->se_client->cl_lock);
>>> + /* If growing the slot table, reset any new sequences to 1 */
>>> + if (highest > ses->se_cb_highest_slot) {
>>> + int i;
>>> +
>>> + for (i = ses->se_cb_highest_slot; i <= highest; ++i) {
>>> + /* beyond the end of the array? */
>>> + if (i >= NFSD_BC_SLOT_TABLE_MAX)
>>> + break;
>> 
>> Nit: why not cap "highest" at NFSD_BC_SLOT_TABLE_MAX before starting
>> this for loop?
>> 
> 
> Duh -- of course. Will fix.
> 
>> 
>>> + ses->se_cb_seq_nr[i] = 1;
>>> + }
>>> + }
>>> + ses->se_cb_highest_slot = highest;
>>> + spin_unlock(&ses->se_client->cl_lock);
>>> +}
>>> +
>>> /*
>>>  * CB_SEQUENCE4resok
>>>  *
>>> @@ -485,7 +518,7 @@ static int decode_cb_sequence4resok(struct xdr_stream *xdr,
>>> p += XDR_QUADLEN(NFS4_MAX_SESSIONID_LEN);
>>> 
>>> dummy = be32_to_cpup(p++);
>>> - if (dummy != session->se_cb_seq_nr) {
>>> + if (dummy != session->se_cb_seq_nr[cb->cb_held_slot]) {
>> 
>> Nit: Let's rename "dummy" as "seqid", and add a "highest" variable
>> for the next XDR field (not shown here).
>> 
> 
> Ok.
> 
>> 
>>> dprintk("NFS: %s Invalid sequence number\n", __func__);
>>> goto out;
>>> }
>>> @@ -496,9 +529,15 @@ static int decode_cb_sequence4resok(struct xdr_stream *xdr,
>>> goto out;
>>> }
>>> 
>>> - /*
>>> -  * FIXME: process highest slotid and target highest slotid
>>> -  */
>>> + p++; // ignore current highest slot value
>>> +
>>> + dummy = be32_to_cpup(p++);
>> 
>> Nit: I prefer a name for this argument variable like "target".
>> 
>> The compiler should be able to combine the usage of these variables
>> into a single memory location.
>> 
> 
> Ok.
> 
>> 
>>> + if (dummy == 0) {
>>> + dprintk("NFS: %s Invalid target highest slotid\n", __func__);
>>> + goto out;
>>> + }
>>> +
>>> + update_cb_slot_table(session, dummy);
>>> status = 0;
>>> out:
>>> cb->cb_seq_status = status;
>>> @@ -1208,31 +1247,38 @@ void nfsd4_change_callback(struct nfs4_client *clp, struct nfs4_cb_conn *conn)
>>>  * If the slot is available, then mark it busy.  Otherwise, set the
>>>  * thread for sleeping on the callback RPC wait queue.
>>>  */
>>> -static bool nfsd41_cb_get_slot(struct nfsd4_callback *cb, struct rpc_task *task)
>>> +static void nfsd41_cb_get_slot(struct nfsd4_callback *cb, struct rpc_task *task)
>>> {
>>> struct nfs4_client *clp = cb->cb_clp;
>>> + struct nfsd4_session *ses = clp->cl_cb_session;
>>> + int idx;
>>> 
>>> - if (!cb->cb_holds_slot &&
>>> -     test_and_set_bit(0, &clp->cl_cb_slot_busy) != 0) {
>>> + if (cb->cb_held_slot >= 0)
>>> + return;
>>> +retry:
>>> + spin_lock(&clp->cl_lock);
>>> + idx = ffs(ses->se_cb_slot_avail) - 1;
>>> + if (idx < 0 || idx > ses->se_cb_highest_slot) {
>>> + spin_unlock(&clp->cl_lock);
>>> rpc_sleep_on(&clp->cl_cb_waitq, task, NULL);
>>> - /* Race breaker */
>>> - if (test_and_set_bit(0, &clp->cl_cb_slot_busy) != 0) {
>>> - dprintk("%s slot is busy\n", __func__);
>>> - return false;
>>> - }
>>> - rpc_wake_up_queued_task(&clp->cl_cb_waitq, task);
>>> + goto retry;
>>> }
>>> - cb->cb_holds_slot = true;
>>> - return true;
>>> + /* clear the bit for the slot */
>>> + ses->se_cb_slot_avail &= ~BIT(idx);
>>> + spin_unlock(&clp->cl_lock);
>>> + cb->cb_held_slot = idx;
>>> }
>>> 
>>> static void nfsd41_cb_release_slot(struct nfsd4_callback *cb)
>>> {
>>> struct nfs4_client *clp = cb->cb_clp;
>>> + struct nfsd4_session *ses = clp->cl_cb_session;
>>> 
>>> - if (cb->cb_holds_slot) {
>>> - cb->cb_holds_slot = false;
>>> - clear_bit(0, &clp->cl_cb_slot_busy);
>>> + if (cb->cb_held_slot >= 0) {
>>> + spin_lock(&clp->cl_lock);
>>> + ses->se_cb_slot_avail |= BIT(cb->cb_held_slot);
>>> + spin_unlock(&clp->cl_lock);
>>> + cb->cb_held_slot = -1;
>>> rpc_wake_up_next(&clp->cl_cb_waitq);
>>> }
>>> }
>>> @@ -1265,8 +1311,8 @@ static void nfsd4_cb_prepare(struct rpc_task *task, void *calldata)
>> 
>> Nit: This patch should update the documenting comment before
>> nfsd4_cb_prepare() -- the patch implements "multiple slots".
>> 
> 
> +1
> 
>> 
>>> trace_nfsd_cb_rpc_prepare(clp);
>>> cb->cb_seq_status = 1;
>>> cb->cb_status = 0;
>>> - if (minorversion && !nfsd41_cb_get_slot(cb, task))
>>> - return;
>>> + if (minorversion)
>>> + nfsd41_cb_get_slot(cb, task);
>>> rpc_call_start(task);
>>> }
>>> 
>>> @@ -1292,7 +1338,7 @@ static bool nfsd4_cb_sequence_done(struct rpc_task *task, struct nfsd4_callback
>>> return true;
>>> }
>>> 
>>> - if (!cb->cb_holds_slot)
>>> + if (cb->cb_held_slot < 0)
>>> goto need_restart;
>>> 
>>> /* This is the operation status code for CB_SEQUENCE */
>>> @@ -1306,10 +1352,10 @@ static bool nfsd4_cb_sequence_done(struct rpc_task *task, struct nfsd4_callback
>>>  * If CB_SEQUENCE returns an error, then the state of the slot
>>>  * (sequence ID, cached reply) MUST NOT change.
>>>  */
>>> - ++session->se_cb_seq_nr;
>>> + ++session->se_cb_seq_nr[cb->cb_held_slot];
>>> break;
>>> case -ESERVERFAULT:
>>> - ++session->se_cb_seq_nr;
>>> + ++session->se_cb_seq_nr[cb->cb_held_slot];
>>> nfsd4_mark_cb_fault(cb->cb_clp);
>>> ret = false;
>>> break;
>>> @@ -1335,17 +1381,16 @@ static bool nfsd4_cb_sequence_done(struct rpc_task *task, struct nfsd4_callback
>>> case -NFS4ERR_BADSLOT:
>>> goto retry_nowait;
>>> case -NFS4ERR_SEQ_MISORDERED:
>>> - if (session->se_cb_seq_nr != 1) {
>>> - session->se_cb_seq_nr = 1;
>>> + if (session->se_cb_seq_nr[cb->cb_held_slot] != 1) {
>>> + session->se_cb_seq_nr[cb->cb_held_slot] = 1;
>>> goto retry_nowait;
>>> }
>>> break;
>>> default:
>>> nfsd4_mark_cb_fault(cb->cb_clp);
>>> }
>>> - nfsd41_cb_release_slot(cb);
>>> -
>>> trace_nfsd_cb_free_slot(task, cb);
>>> + nfsd41_cb_release_slot(cb);
>>> 
>>> if (RPC_SIGNALLED(task))
>>> goto need_restart;
>>> @@ -1565,7 +1610,7 @@ void nfsd4_init_cb(struct nfsd4_callback *cb, struct nfs4_client *clp,
>>> INIT_WORK(&cb->cb_work, nfsd4_run_cb_work);
>>> cb->cb_status = 0;
>>> cb->cb_need_restart = false;
>>> - cb->cb_holds_slot = false;
>>> + cb->cb_held_slot = -1;
>>> }
>>> 
>>> /**
>>> diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
>>> index 5b718b349396f1aecd0ad4c63b2f43342841bbd4..20a0d40202e40eed1c84d5d6c0a85b908804a6ba 100644
>>> --- a/fs/nfsd/nfs4state.c
>>> +++ b/fs/nfsd/nfs4state.c
>>> @@ -2002,6 +2002,8 @@ static struct nfsd4_session *alloc_session(struct nfsd4_channel_attrs *fattrs,
>>> }
>>> 
>>> memcpy(&new->se_fchannel, fattrs, sizeof(struct nfsd4_channel_attrs));
>>> + new->se_cb_slot_avail = ~0U;
>>> + new->se_cb_highest_slot = battrs->maxreqs - 1;
>>> return new;
>>> out_free:
>>> while (i--)
>>> @@ -2132,7 +2134,9 @@ static void init_session(struct svc_rqst *rqstp, struct nfsd4_session *new, stru
>>> 
>>> INIT_LIST_HEAD(&new->se_conns);
>>> 
>>> - new->se_cb_seq_nr = 1;
>>> + for (idx = 0; idx < NFSD_BC_SLOT_TABLE_MAX; ++idx)
>>> + new->se_cb_seq_nr[idx] = 1;
>>> +
>>> new->se_flags = cses->flags;
>>> new->se_cb_prog = cses->callback_prog;
>>> new->se_cb_sec = cses->cb_sec;
>>> @@ -3159,7 +3163,6 @@ static struct nfs4_client *create_client(struct xdr_netobj name,
>>> kref_init(&clp->cl_nfsdfs.cl_ref);
>>> nfsd4_init_cb(&clp->cl_cb_null, clp, NULL, NFSPROC4_CLNT_CB_NULL);
>>> clp->cl_time = ktime_get_boottime_seconds();
>>> - clear_bit(0, &clp->cl_cb_slot_busy);
>>> copy_verf(clp, verf);
>>> memcpy(&clp->cl_addr, sa, sizeof(struct sockaddr_storage));
>>> clp->cl_cb_session = NULL;
>>> diff --git a/fs/nfsd/state.h b/fs/nfsd/state.h
>>> index 41cda86fea1f6166a0fd0215d3d458c93ced3e6a..2987c362bdd56251e736879dc89302ada2259be8 100644
>>> --- a/fs/nfsd/state.h
>>> +++ b/fs/nfsd/state.h
>>> @@ -71,8 +71,8 @@ struct nfsd4_callback {
>>> struct work_struct cb_work;
>>> int cb_seq_status;
>>> int cb_status;
>>> + int cb_held_slot;
>>> bool cb_need_restart;
>>> - bool cb_holds_slot;
>>> };
>>> 
>>> struct nfsd4_callback_ops {
>>> @@ -307,6 +307,9 @@ struct nfsd4_conn {
>>> unsigned char cn_flags;
>>> };
>>> 
>>> +/* Max number of slots that the server will use in the backchannel */
>>> +#define NFSD_BC_SLOT_TABLE_MAX 32
>>> +
>> 
>> The new comment is unclear about whether this is an implementation
>> limit or a protocol limit. I suggest:
>> 
>> /* Maximum number of slots that NFSD implements for NFSv4.1+ backchannel */
>> 
>> And make this "sizeof(u32) * 8" or something similar that documents
>> where the value of this limit comes from.
>> 
> 
> Ok.
> 
>>> /*
>>>  * Representation of a v4.1+ session. These are refcounted in a similar fashion
>>>  * to the nfs4_client. References are only taken when the server is actively
>>> @@ -325,7 +328,9 @@ struct nfsd4_session {
>>> struct nfsd4_cb_sec se_cb_sec;
>>> struct list_head se_conns;
>>> u32 se_cb_prog;
>>> - u32 se_cb_seq_nr;
>>> + u32 se_cb_slot_avail; /* bitmap of available slots */
>>> + u32 se_cb_highest_slot; /* highest slot client wants */
>>> + u32 se_cb_seq_nr[NFSD_BC_SLOT_TABLE_MAX];
>>> struct nfsd4_slot *se_slots[]; /* forward channel slots */
>>> };
>>> 
>>> @@ -459,9 +464,6 @@ struct nfs4_client {
>>>  */
>>> struct dentry *cl_nfsd_info_dentry;
>>> 
>>> - /* for nfs41 callbacks */
>>> - /* We currently support a single back channel with a single slot */
>>> - unsigned long cl_cb_slot_busy;
>>> struct rpc_wait_queue cl_cb_waitq; /* backchannel callers may */
>>> /* wait here for slots */
>>> struct net *net;
>>> diff --git a/fs/nfsd/trace.h b/fs/nfsd/trace.h
>>> index f318898cfc31614b5a84a4867e18c2b3a07122c9..a9c17186b6892f1df8d7f7b90e250c2913ab23fe 100644
>>> --- a/fs/nfsd/trace.h
>>> +++ b/fs/nfsd/trace.h
>>> @@ -1697,7 +1697,7 @@ TRACE_EVENT(nfsd_cb_free_slot,
>>> __entry->cl_id = sid->clientid.cl_id;
>>> __entry->seqno = sid->sequence;
>>> __entry->reserved = sid->reserved;
>>> - __entry->slot_seqno = session->se_cb_seq_nr;
>>> + __entry->slot_seqno = session->se_cb_seq_nr[cb->cb_held_slot];
>>> ),
>>> TP_printk(SUNRPC_TRACE_TASK_SPECIFIER
>>> " sessionid=%08x:%08x:%08x:%08x new slot seqno=%u",
>>> 
>>> -- 
>>> 2.47.0
>>> 
>> 
> 
> -- 
> Jeff Layton <jlayton@...nel.org>


--
Chuck Lever

Powered by blists - more mailing lists